Stock AI Voice, Real Voice Actor: Why You're Liable, Not the Provider

What's at stake

If you pick a ready-made AI voice from a catalog for an ad, a podcast or a YouTube channel, you probably assume something simple: the provider checked the voice, so I'm safe. Two rulings now show that assumption doesn't hold. If the catalog voice sounds like a real person, whoever publishes it is liable. The provider isn't. We compare how different countries protect voices, what the platforms promise in their terms and what that means for you in practice.

Warsaw: a designer, a library, a ruling

On November 13, 2026, Judge Sylwia Wit vel Wilk of the Warsaw District Court's intellectual property division handed down Poland's first ruling on an AI voice. The plaintiff was voice actor Jarosław Łukomski. The company JFC Polska had advertised a home sewage treatment plant using a voice that, according to a court-appointed expert, sounded “clearly, deceptively similar” to his for the average listener. According to Wirtualnemedia, the spots ran in Polish, Czech, Slovak and Hungarian. OKO.press reports that acquaintances sent Łukomski the link because they thought he was doing the ads himself.

One piece of testimony is key. The JFC Polska graphic designer in charge said he used ElevenLabs' tool, picked a voice from its library of available AI voices and only changed the pace and emotional tone. The managing director told OKO.press:

“We were convinced we were acting lawfully.”

That didn't help. The court classified the spots as a deepfake. What matters, it said, is whether listeners could take the AI statement for real. Every person's voice is a protected personal interest (dobro osobiste), regardless of how well known they are. You don't need a special “premium voice” for that protection. JFC Polska must remove the material, stop using it and publicly apologize: for 45 days on Facebook and 14 days on its own website and on Wirtualnemedia. On top of that comes 30,000 złoty in damages plus interest and just over 17,000 złoty in compensation. The claim was for 50,000 złoty in damages and 102,000 złoty for unjust enrichment. The court had doubts about how the losses were calculated. The defendant had submitted a 422-page private expert report arguing that similar sound doesn't prove copying. The ruling isn't final.

The case didn't establish how the voice model was created at ElevenLabs. Tomasz Bartos of the voice talent database Mikrofonika, which supports Łukomski, announced legal action against ElevenLabs and is demanding to know how the platform handles voices. We found no statement from ElevenLabs.

Berlin 2025: “Irrelevant whether the defendant paid”

Berlin Regional Court II (LG Berlin II) reached a similar decision back on August 20, 2025 (case no. 2 O 202/24). A YouTuber had used an AI imitation of Manfred Lehmann, the German dubbing voice of Bruce Willis, in two videos. According to LTO, his defense was that he had just been looking for a suitable voice and took the one the software suggested. According to PPC Land, the court found it irrelevant whether the defendant had paid the AI provider and acquired a usage right. The provider's consent doesn't replace the consent of the person whose voice is being imitated.

Lehmann didn't have to prove confusion in the strict sense. The court looked at whether a not insignificant part of the audience could believe Lehmann had spoken himself. Comments under the videos in which viewers recognized his voice served as evidence. The court counted the missing AI disclosure as an aggravating factor. It awarded a notional license fee of €2,000 per video, so €4,000 in total, plus legal costs. According to dejure.org, an appeal is pending at the Kammergericht, Berlin's higher regional court (10 U 100/25).

Four legal systems compared

Two things stand out. First, in Europe it's users who pay, not platforms. In both Poland and Berlin, only the user was sued. The Danish bill L 49, introduced by Culture Minister Zenia Stampe on October 8, 2026, reverses the burden of proof: anyone who publishes a realistic replica of someone's voice, appearance or movement must prove consent. According to aimusicpreneur.com, that applies even to accidental uploads. Hosting services as defined in the Digital Services Act are exempt, and the bill doesn't regulate tool providers. Satire and parody remain allowed. Protection is meant to last until 50 years after death, and the law is planned to take effect on January 1, 2027. L 49 hasn't been passed yet.

Second, only the US goes after the tools. The ELVIS Act also covers providers whose technology is primarily designed to replicate voices without permission. A general-purpose generator with a voice library probably doesn't fall under it. The NO FAKES Act (S. 4591) cleared the Senate Judiciary Committee in June 2026, but a floor vote is still pending. If the session ends on January 3, 2027, without a vote, the bill has to be reintroduced. According to humva.com, an AI disclaimer would explicitly not be a defense under it. California regulates contracts for digital replicas with AB 2602 and replicas of deceased performers with AB 1836.

Protection is tightening elsewhere too. On June 24, 2026, France's Cour de cassation (no. 25-20.483) recognized a right to one's own voice on a par with the right to one's own image, though it requires balancing it against artistic freedom. In the case of actor Denis Podalydès, whose cloned voice read philosophy lessons on a YouTube channel, Google was ordered in summary proceedings to hand over the channel operator's data, according to the law firm Kohen Avocats. In Shanghai, after an appeal ruling on September 29, 2026, a company now has to prove where a similar-sounding AI voice came from. The excuse that the employee responsible had left didn't cut it.

What providers promise and what they don't

The ElevenLabs terms of service even state explicitly that two parties may receive “the same or similar output” and that you must review the output before using it. The usage policy prohibits impersonating living people and offers a separate form for reporting voices used without permission.

Our take: the verification checks whether the person uploading a voice matches the voice in the training recordings. From everything publicly described, it doesn't check whether the result sounds, by accident or by design, like a well-known voice actor. A good voice impersonator could clear that bar legally. That doesn't protect the voice actor being imitated, and it doesn't protect you as a user either.

Does mandatory labeling help?

Since August 2, 2026, Article 50(4) of the AI Act has required deployers to clearly label deepfakes. That explicitly includes audio resembling existing people. Violations can bring fines of up to €15 million or 3 percent of revenue. But the obligation gives the voice actor no claim and gives you no permission. The Berlin court did count the missing disclosure as an aggravating factor. A disclosure wouldn't have cured the infringement, though, because the voice was still being used to drive reach.

The dispute with Netflix shows how live this issue is in Germany. According to Netzpiloten, almost 80 percent of German dubbing actors are refusing to sign an AI training clause. A legal opinion commissioned by the Verband Deutscher Sprecher, the German voice actors' association, considers the clause invalid. According to the Berliner Kurier, Matti Klemm, the German voice of Jason Momoa, is suing Netflix.

How to use stock AI voices without nasty surprises

1. Run a blind test. Play the voice to three people and ask: “Who does this remind you of?” That's exactly the impression courts look at in the average listener. Well-known ad voices, dubbing actors and audiobook narrators are the biggest risks.
2. Don't fine-tune toward someone. Never push pace, pitch or emotion toward a famous voice, and never label anything “like Bruce Willis.”
3. For ads, only with an explicit license. Either sign a direct contract with the voice actor or use a licensed offering like ElevenLabs' “Iconic” catalog, where the license is negotiated separately. Under L 49, you need to be able to prove consent.
4. Document everything. Record the voice ID, profile page, date, plan and the version of the terms of service (addendum dated March 6, 2026).
5. Label it. In the EU, disclose that it's an AI voice. It's mandatory, but it doesn't replace permission.
6. Check every country. JFC Polska ran its spot in four languages. Every market comes with its own laws.
7. Agencies: settle the risk by contract. Agree with your client who's responsible for choosing the voice. According to its terms, the provider won't take that on.
8. Stop immediately if you get a cease-and-desist letter. Take the content offline and get legal advice. JFC Polska pulled the spots after the first demand and was still found liable.

Our verdict

Poland, Germany, France and soon Denmark take different routes to the same result: what counts is whether it sounds like a real person is speaking. Where the voice came from doesn't matter. You don't need to be famous to be protected. Fame does raise the amount, though: Lehmann got the standard fee, Łukomski about a third of what he had demanded. The providers have shifted the risk entirely onto users in their terms. Our take: until lawmakers go after the tools, as Tennessee has, a blind test before publishing is your best protection. Open models like Sophon-Realtime or ZONOS3, which clone voices from a few seconds of audio, make that check even more important. Whether Mikrofonika's lawsuit against ElevenLabs changes anything remains to be seen. Neither ruling is final yet.